The following article was written by Jason Liu (@jxnlco) and originally published as a thread on X on May 20, 2026. It is read here in full with the author's permission for the AI Daily Briefing podcast.

CODEX: FROM CODING ASSISTANT TO COMPUTER-WORK SYSTEM
by Jason Liu

Codex started with coding tasks. Inspect the repo, make a diff, run tests, open a PR. That's the entry point most people see. But most real computer work is already code-mediated. Shell commands, web browsing, APIs, documents, spreadsheets, automations — these are all things you can express through code. So the natural extension is for Codex to stop being a coding assistant and start being a computer-work system.

The Codex app makes this concrete. Here's what's under the hood.

Durable Threads. These are long-running threads that keep working context across sessions. You pin the ones you come back to constantly — Chief of Staff, release process, documentation review, external monitoring — and use Command-1 through Command-9 to jump to them. They're persistent workspaces, not short chats. The thread remembers.

Voice Input. Great for rough ideas or dumping two to three minutes of half-formed thoughts. Voice works especially well with transcripts. Raw meeting notes are often better than polished summaries because they carry the uncertainty and the unfinished edges. You can say "someone named Ben mentioned something about this in Slack" and Codex figures it out.

Steering versus Queuing. These are two different interaction patterns. Steering means interrupting a task that's in progress — "make this smaller," "fix the spacing," "stop and show me what you have." Queuing means adding the next task after the current one finishes — "once that's done, send the preview link to Slack." Steering is for course correction. Queuing is for batching intent.

Tools and Reach. Codex has a growing set of tools. There's the dollar-sign browser, which is an in-app side-panel browser for inspecting and annotating. There's at-chrome, which gives access to your signed-in Chrome workflows. At-computer handles desktop GUI tasks. Then there are MCP servers and connectors for Slack, Gmail, and Calendar. And Skills — reusable packaged workflows that you or your team create.

Work from Anywhere. The mobile app lets you start a task on desktop, then continue, steer, or approve from your phone. Your local environment stays intact on the machine where it's running. You're not remote-controlling a cloud VM. Your actual computer does the work. You just check in from anywhere.

Automations. There are two kinds. Scheduled automations start fresh on a schedule — a daily report, a repo health check, a morning summary. Thread automations are different. They wake up the same thread on a schedule. Think heartbeat-style — every thirty minutes, check Slack and Gmail, draft replies, surface what needs attention. This is perfect for feedback loops: PR comments, Google Doc comments, Slack threads you've been mentioned in.

Goals. Goals are long-running tasks with a clear finish line and a verifier. The example Jason gives: migrate a tool from Python to Rust, and keep going until all unit tests pass. The verifier can be a test suite, a benchmark, a bug reproduction script — something that objectively proves the work is done. Goals don't stop until the verifier says stop.

Side Panel. Keep artifacts right beside the thread. Code, slide decks, PDFs, browser pages, tables, data apps. The side panel supports inspection, annotation, operating web surfaces, and reviewing changes. It's especially good for things like index.html previews, Storybook, Remotion Studio, browser-based slide decks, and data applications. You review and edit in place without context switching.

Shared Memory and the Vault Pattern. This is how you give Codex durable context that survives individual threads. Use an Obsidian vault — or any folder structure with plain files. The typical layout has a TODO.md, a people directory, a projects directory, an agent directory, and a notes directory. Add an AGENTS.md file that tells Codex how to maintain the vault. First-party Memories and Chronicle features also help, but the vault pattern is the anchor.

The bottom line. Codex is shifting from coding assistant to general computer-work system. The real power comes from combining durable threads, tools, automations, Goals, the side panel for review, and shared memory for persistence. Each piece is useful alone. Together they form a system where the computer does the work, and you steer it.

This article was written by Jason Liu and originally published as a thread on X at twitter dot com slash jxnlco. Thank you for listening to this special episode of the AI Daily Briefing.